Extrahering av täta funktioner för segmentering på urban vägmiljö

Modern infrastructure relies heavily on the quality of road surfaces. The industry responsible for maintaining these assets remains extensive yet still largely dependent on manual inspection. Automating this process and improving the quality of information available to stakeholders could significantly reduce both operational costs and environmental impact. Since road condition is typically assessed visually, computer vision techniques offer a promising approach. Within this field, a particularly challenging task is the automatic segmentation of road surface damage, due to the small spatial extent and subtle appearance of most defects. Recent advances in computer vision have produced general-purpose feature extractors capable of solving a wide range of visual tasks. These models, trained on large and diverse datasets, have shown strong performance in broad domains. This thesis explores whether such general models can also be effective in narrow, specialized tasks such as road damage segmentation. The fine-grained and imbalanced nature of the target classes in this setting makes it non-trivial to use features extracted from general feature extractors. To further improve performance on dense prediction tasks, post-pretraining, a self-supervised adaptation step has been proposed to refine general models. While this technique has shown promise in general domains, its applicability to domain-specific problems like road damage segmentation remains largely unexplored. This thesis investigates two related questions: (1) How can a general feature extractor be most effectively applied to road damage segmentation? and (2) Can its performance be improved through post-pretraining? To answer the first question, we evaluate several decoder architectures that map general features to dense segmentation outputs. For the second, we assess whether post-pretraining on general and domain-specific datasets improves model performance in the road domain. Our findings show that general feature extractors, when paired with convolutional decoders, can perform well in this setting. However, the benefits of post-pretraining observed in broad domains do not transfer to road-specific data.